Axiom Math scored 120/120 on the 2025 Putnam exam, beating the best LLM (DeepSeek at 103/120) and the best human contestant (110/120).

“The best LM deepseek got 103 points out of a 120 point exam. The best human obviously we now know is a student from either MIT or Chicago. We don't know which one because they don't announce the top five winner score got 110 and we got 120.”

GRPO became dominant not due to algorithmic novelty but because DeepSeek did the scaling engineering and shipped a working artifact model that validated it.

“I think the reason GRPO specifically like that algorithm and that acronym like, you know, very concretely took off was not necessarily because it was like a big quantum leap on what came before. It was because DeepSeek did a lot of engineering work around actually scaling it and released an actual artifact model that worked really well with it.”

DeepSeek delivered GPT-4-level performance at 1/600th the inference cost of GPT-4.

“Deep Seek, for example, on GPT-4 was 1/600 the cost.”

DeepSeek uses expert parallelism rather than pipeline parallelism for frontier-scale inference, a deployment choice worth tracking as MoE models proliferate.

“The DeepSeek paper reports what they do, which is that they just do a lot of expert parallelism.”

Banning DeepSeek in the US would likely accelerate global adoption of DeepSeek models and Huawei chips outside American borders.

“If we ban it boy it's just going to one you know even more people working with Huawei chip and deepseek trying to optimize it. But two, you know, in this world we've created where there's so much animosity, I think people would be glad to use it in these other countries.”

Brad Carson contends that Chinese AI firms like DeepSeek are not simply extensions of the Chinese Communist Party, pushing back on a common Western assumption.

“These companies are not co-terminus with Chinese Communist Party.”

DeepSeek cited as the key mechanism by which China could close the AI compute gap with the US within 6 months if US doesn't expand capacity.

“If we don't build more compute capacity in the next 6 months, the Chinese will catch up with us with deepseek.”

Manus, an AI agent that has operated exclusively on US AI models rather than DeepSeek or other Chinese models, worth watching as a bellwether for AI stack allegiance.

“They've only operated on US models.”

DeepSeek (version 4 or equivalent) is flagged as a near-term open-weight release to watch, potentially carrying a major architectural change and historically timed around the Chinese New Year.

“Historically around the Chinese New Year, there have been always a lot of model releases, openw weight model releases. So maybe there's like a little surprise in there. Maybe we'll see Deepseek version 4 and maybe there is a bigger change.”